Deep Learning Identifies Fragile X Syndrome EEG Biomarkers

A novel deep learning framework has been developed to automatically identify electroencephalography (EEG) biomarkers for Fragile X Syndrome (FXS), a neurodevelopmental disorder. The framework integrates multiple representation types, including convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and recurrence plot (RP) analysis. This approach aims to capture the complex spatial, temporal, and nonlinear dependencies within EEG signals. The method involves decomposing band-limited EEG signals into alpha and gamma components, which are then transformed into various representations such as temporal feature sequences, time-frequency maps, and RP images. CNN modules are designed to learn discriminative spatial-spectral and dynamical textures from image-based representations, while LSTM modules model the temporal modulation of oscillatory activity. Subject-independent evaluation demonstrated that this hybrid CNN-LSTM architecture significantly outperforms single-modality baselines. Gamma features showed strong discriminative power, and the integration of both alpha and gamma features yielded the best overall performance. These findings suggest that deep learning with nonlinear representations offers a scalable approach for developing EEG biomarkers for FXS, potentially aiding in diagnosis, stratification, and treatment monitoring in clinical settings.